# Out of scope and wishlist
Two lists. **Out of scope** holds things `trg` deliberately does not do, each
with the reason — revisit one only when there is a *new* reason, not because it
looks like an obvious gap. **Wishlist** holds things that could reasonably be
added later.
When an item moves into the codebase, delete it here and record it in
[CHANGELOG.md](CHANGELOG.md).
## Out of scope
### Non-goals
- **Matching ripgrep's throughput.** Searching compressed logs is
inflate-bound; the regex engine contributes almost nothing. Only the inflate
implementation moves CPU.
- **Replacing `rg` for uncompressed source trees.** `trg` is for log archives.
- **Forking ripgrep or reimplementing its regex/search engine.** `trg` is a thin
CLI over ripgrep's published crates.
- **Searching `$PWD` when no `PATH` is given.** An unset `$LOGS` would turn into
a confident exit `1` from data nobody named. A missing `PATH` is a usage
error (`2`).
- **A `-z` flag or filename-based format detection.** Compression is detected
from content; extensions lie.
- **Parallelism inside a single gzip archive.** gzip has no index, so member N
cannot be reached without inflating members 1..N-1. This is the format, not
an implementation gap.
### Not built, for now
- **bzip2, lz4.** None exist on the target servers and none are expected.
Each would be mechanically simple (a sniff entry and a decoder) — the reason
to add one is a real archive in that format.
- **`.zip` archives.** Not "another decoder": zip's central directory makes
members individually seekable and skippable, which the streaming
`for_each_member` model cannot express. Skipping is the only reason to want
zip, so it needs a second traversal model. Reconsider if zip archives show up.
## Wishlist
### Features
- **Adaptive `--load-limit`** — a feedback loop that adjusts concurrency during
the run, instead of a one-off check at startup.
- **zstd whole-buffer inflate** — zstd's `Frame_Content_Size` header is a more
trustworthy size hint than gzip's mod-2³² trailer. Low priority: zstd already
decompresses several times faster than gzip, so buffer mode gains least here.
- **Intra-archive parallelism for zstd** — frame boundaries make it possible,
unlike gzip. It reopens `src/sched.rs`, which has produced four deadlocks;
treat it as a project, not a patch.
- **A clearer `-c` signal for match-then-binary members** — stdout stays
rg-compatible today and a stderr note breaks the silence, but a script that
reads only stdout still has to check `$?`.
### Engineering
- **Regression test for sched deadlock #4** (teardown guard armed after the
spawn loop). Needs a fault-injection hook around `scope.spawn`.
- **Retire `--inflate=buffer` entirely?** Since 0.6.1 `auto` streams: with
zlib-ng streaming, the buffered path never used less total CPU on any machine
measured (`bench/inflate_modes.py`, 2026-10-23), and on 42 MB archives used
8-59% more. It survives as an explicit opt-in. Removing it would drop
`--inflate`, `--inflate-budget` and the libdeflater dependency, which breaks
scripts passing those flags. (The first measurement, 1.8-2.7x slower, was
wrong: those archives were 1.4 MB over the 1 GiB budget, so every buffered
row measured a failed attempt plus a full stream.)
- **`-g` without a `/` could match the member's file name**, as
`--archive-glob` and ripgrep do. Today `-g access.log` never matches
`logs/access.log`; only patterns that start with `*` work that way, because
`*` crosses `/`. Changing it changes which members existing globs select.